TriAgentLoop
An open-source setup that splits AI coding work between three agents. None of them is allowed to sign off on its own work.
An agent reviewing its own work is not reviewing anything. So the three roles work on separate branches, and nothing gets merged without a written review from a different one.
Facts
- One agent plans, one builds, one reviews. Each works in its own sandbox on its own branch
- Every handoff between them is written to a log that can only be appended to, so the reasoning survives after the session ends
- Work only gets merged once the reviewing agent has written down that it approves
Stack
- TypeScript
- Rust
- Git
This site was built with it. The plan, the review findings, the disagreements and how they were settled all sit in a log in the repository, so the record of how the work got decided is as readable as the work itself.